A multi-scale channel attention neural network with feature pre-extraction based on ERT in two-phase flow pattern recognition

Abstract To improve the accuracy and efficiency of two-phase flow pattern recognition in electrical resistance tomography (ERT) measurement scenarios and address the limitations of conventional methods in real-time monitoring and large-scale automated processing of reconstructed ERT images, a multi-scale channel attention neural network with feature pre-extraction (PE-MCNN) is proposed. The model is specifically designed for automated flow pattern analysis after ERT image reconstruction, leveraging images generated during the ERT reconstruction process to construct its training set. Through a multi-stage network architecture, PE-MCNN achieves synergistic optimization of feature PE and attention mechanisms, enabling high-throughput recognition of flow patterns from bulk ERT-reconstructed images. For validation, comprehensive evaluations were conducted on four datasets: a labeled image dataset, a generalized dataset, a variable-conductivity dataset featuring gas bubbles with deliberately modulated electrical conductivity to test robustness against permittivity variations common in propellants, and a simulated dataset replicating high-speed turbopump operating conditions, containing 246 experimentally acquired images reconstructed using ERT techniques. Experimental results demonstrate that PE-MCNN outperforms existing methods, attaining an average recognition accuracy exceeding 0.944 across all test sets, with 0.9837 accuracy on the ERT experimental dataset. Notably, the proposed method eliminates the need for manual feature engineering and significantly enhances processing efficiency, achieving fully automated recognition of complex flow patterns at scale. This breakthrough overcomes the characterization limitations of traditional approaches under dynamic industrial conditions and establishes a reliable post-ERT processing framework for real-time visual monitoring in industrial pipeline systems.